Question How to improve PCIe Gen5 performance of AMD 7600X?

[HLGM]

Now I am using a B650E motherboard with 7600X CPU to measure the PCIe Gen5 devices. Unfortunately, I find the PCIe Gen5 slot can not get up to full speed. It seems the PCIe Gen5 x16 slot has only x8 capability.

Using B650E and X670E MB with 7600X CPU to test PCIe Gen5 performance. The configuration is as below.
1. Configure bifurcation in BIOS as x4x4x4x4 of PCIe x16 slot.
2. Connect 4 PM1743 SSDs to the PCIe x16 slots through a x4x4x4x4 PCIe Gen5 retimer.
3. Run IOStress.

The first SSD is connected, I can get about 14GB/s IO throughput. Add the second SSD in system, the throughput can be up to 29GB/s or so. When I add one more SSD again, the IO throughput will still be 29GB/s. Even if I connect 4 SSDs, the IO throughput is also 29GB/s. OS can detect all 4 SSDs and LnkStas are all 32G x4. So it looks the 7600X PCIe Gen5 x16 slot has only a x8 capability.

I am not sure how method can help to improve the PCIe slot up to real x16 capability?

 

[Tech Junky]

If you have two devices it cuts the bandwidth to x8/x8 in most cases. Look at the specs for the mobo. Also, the B series won't have as many lanes as the x670e would.

 

[SteinFG]

Try asking on Level1Forums. My guess is that x4x4x4x4 is not officially supported, so there may be bugs.

 

[HLGM]

If you have two devices it cuts the bandwidth to x8/x8 in most cases. Look at the specs for the mobo. Also, the B series won't have as many lanes as the x670e would.

Thanks a million. The configuration x4x4x4x4 is used to assign dedicated 4lanes to individual SSD. In theory, 4 SSDs must show us the performance more than 50GB/s in sequential reading test. But it is not true.
I have also tried the "ROG CROSSHAIR X670E GENE" from ASUS, of course, CPU is still 7600X. The same phenomena appears. The IO throughput is locked on to x8 instead of x16.

 

[Tech Junky]

The math says ~64GB/s on the slot. Take away overhead and divide by 4 and maybe get 15 per device.

Now, the question is how does the CPU handle that kind of bandwidth when switching between drives.

If there's nothing in any other slots like a GPU then it shouldn't lock at x8 and split that into quarters.

One thing that's comes to mind is the M2 slots 1&2 are Gen 5 x4 each and if you're using either, then it might down clock the pcie slot. I've seen some weird allocations with some boards I've looked into for potentially jumping to AMD.

 

[HLGM]

Try asking on Level1Forums. My guess is that x4x4x4x4 is not officially supported, so there may be bugs.

Thanks, buddy.

 

[Gheis]

Your benchmark software is CPU/memory bottlenecked, because 29 GB/s looks a lot like single thread memory benchmark.

 

[JoeRambo]

Your benchmark software is CPU/memory bottlenecked, because 29 GB/s looks a lot like single thread memory benchmark.

An easy test would be 7900X or 7950X. 29GB/S of disk I/O out of 6C consumer part is nice, but would not make much sense in real world as usually the ratio of IO bw to CPU processing is quite higher than dry bw test requires.
IF the new Threadrippers are PCIE5, i'd look into them to have real I/O machine.

 

[HLGM]

Thanks so much all. I have solved the problem. Insert another DIMM and get improved the PCIe Gen5 IO performance. At the begining, we are using single DIMM in system, the IO is limitted. Now the 2nd DIMM is inserted, the issue is solved.

 

[biostud]

Thanks so much all. I have solved the problem. Insert another DIMM and get improved the PCIe Gen5 IO performance. At the begining, we are using single DIMM in system, the IO is limitted. Now the 2nd DIMM is inserted, the issue is solved.

So the lesson here is that posting full system spec when troubleshooting is a good idea.